Hacker News Reader

Comment by braindead_in | Hacker News Reader

We are an manually powered audio transcription service. We convert spoken audio files to text via a four step process which guarantees minimum 98% accuracy. Our customers are mostly researchers, journalists, videographers, podcasters etc.

Right now, our transcribers manually do the typing. If we have a ASR based system, which produces even a 80% accurate transcript of the audio, we can improve our efficiency manifold.

We have a lot of data, audio files and their corresponding transcript. That data can be used to train a DNN based system. A software solution or framework which can do that would be ideal for us so that we can keep feeding in the data and continuously improve our model.

View on HN

ASR will not help improve efficiency unless you don't care about the mistakes made by ASR (for example, close captioning of real-time audio). Why? Because transcribers will need to listen to audio again to be able to correct the mistakes made by ASR. Basically, once ASR listening to audio and then transcriber listening the audio again to correct mistakes made by ASR. Also, you will increase you cost as you will need costly "editors" to correct the ASR mistakes instead of cheaper "transcribers" to transcribe, reviewing and transcribing requires different skill set.

At one time I invested in a medical transcription business that experimented with voice-to-text technologies but couldn't generate any additional efficiency in transcription by using technology. What worked in improving efficiency was the transcriber's knowledge of medical terminology and medical field, matching the same transcriber with the same speaker as transcriber became acclimated with the nuances and tones of the speaker, and letting transcriber handle only certain type of transcription.

Transcription is a manual operation and Toyota Way is a good way to maximize efficiency.

I'll counter that ASR does improve the efficiency. The biggest player in Medical Transcription is Nuance. It's main product is a ASR! The MT industry has moved on from all human process in the last decade. The ASR accuracy in MT is currently around 90%. The rest of the work is done by human editors.

Our goal with ASR is not replace the typists, but to assist them. On an average typing takes 4X the duration of actual time. An 80% accurate ASR may reduce it to 2X, which in turn should translate to faster turnaround and improved throughput.

End-to-end speech recognition with deep recurrent neural networks is currently the state of the art.

See https://github.com/pannous/caffe-speech-recognition (still under development)

Interesting and thanks for your elaboration

Reply on news.ycombinator.com